The specialized unit will work with the provider to organize games in which humans will have to converse with other players and guess whether they are on their team or not.

The government has commissioned a project in which board games will test humans’ response to “anthropomorphic” artificial intelligence systems designed to be deceptive.

The initiative, which ultimately aims to determine how such deception could be used for more nefarious purposes, will be overseen by the AI ​​Safety Institute (AISI). This specialized unit was created a year ago within the Department of Science, Innovation and Technology with a mission to study the security implications of new “frontier” AI tools, a classification that includes OpenAI’s ChatGPT, Google Gemini – formerly Bard – and the Llama 2 model developed by Facebook-owner Meta.

AISI’s latest work program focuses on “exploring the limits of the ability of border AI systems to persuade, deceive or manipulate people”, according to the text of a contract between the government body and the American technology company Scale AI, which supports the realization of the project.

“For this project, we are specifically interested in the ability of AI systems to engage in anthropomorphic behavior by studying variants of high-performing models that are incentivized to act as humanly as possible,” the contract adds. “This research is designed to explore human resilience in the face of harmful use cases, such as the deployment of advanced AI for disinformation or financial fraud campaigns. The ultimate aim is for Her Majesty’s Government to better understand the capabilities of border AI agents to: persuade and deceive human participants; and convincingly engage in anthropomorphic behavior.

To achieve such an understanding, the project will make different versions of a multiplayer game featuring two teams, divided along binary demographic lines. If eight players participate, four each will come from the United States and the United Kingdom respectively, the contract specifies. Other 50/50 splits likely include groups of women versus men, and under 30 versus over 50.

Individual players in these groups, whether human or AI, will score or lose points by guessing – based on conversational responses – whether or not other participants are on their team.

After completing the first stage of the project, in which “anthropomorphism prompts” will be designed for AI simulations, the department will then conduct second stage studies in which the game is played only by humans. Ultimately, 400 people will participate in 50 matches of eight players each.

These will be followed by a third stage comprising 50 more games – but this time, each featuring four humans and four AI players.

The final step to be completed during the initial three-month engagement will include “data analysis and reporting, carried out internally by AISI,” the contract adds.

Related content

If the institute chooses to extend the agreement for two more months, work could then return to step one with renewed efforts to “rethink anthropomorphism prompts” for possible further studies.

The business deal explains that these designs will be “tailor-made anthropomorphic prompts (that) adapt versions of existing Great Frontier language models, (and) resemble humans as closely as possible and attempt to convince their interlocutor that they are human “.

The document adds: “We would like the vendor to create eight “characters” of the model, which tell the AI ​​which character it should play and how it should behave in different situations. The prompt should also specify the AI ​​player’s goals in the context of the multiplayer deception game.

The contract further explains that the project’s intention is to replicate the type of circumstances in which people could be fooled by generative AI tools – without any real risk.

“The challenge is to measure the psychological capabilities of AI systems to deceive or manipulate without: exposing participants to a real risk of harm; or create a situation in which AISI is required to manage compromising personal data,” it says. “To do this, the study will be designed so that participants will not disclose any personally identifiable information.”

The supplier will be paid £43,462 to fulfill the initial three-month contract. San Francisco-based Scale AI says its products were developed for use by machine learning engineers and can “rapidly deliver large volumes of unbiased and highly accurate training data.”

The initiative to study the deceptive capabilities of artificial intelligence comes just weeks after it was also revealed that AISI had signed a deal worth almost £500,000 to carry out a project to replicate possible cyberattacks in which bad actors could using generative AI to help create chemical weapons or disrupt democratic processes.